AI for industrial control: hocus pocus or real outcomes?

In recent years, artificial intelligence and machine learning have been touted as a panacea for process plants: using instrumentation to gather the data and automated systems to effect improvements to the operations. There have been huge advances and some great successes, but maintaining a healthy cynicism is essential.

Machine learning is certainly one of the most exciting technologies in recent times, and there are three major types of machine learning: supervised, unsupervised and reinforcement learning.

Supervised learning is based on the relationship of given inputs to a set of outputs. For example, the sensor inputs in a process plant (using previous breakdown knowledge) are ‘wired’ to predict the likelihood of the failure of a pump, or a maintenance prompt for equipment.

Unsupervised learning is where no predefined relationship exists between input data and an output variable. For example, a sensor is programmed to log data from hundreds of similar industrial plants, searching for specific patterns for optimal operations. The algorithm is able to trawl large data sets and then classify the data.

Reinforcement learning uses algorithms to perform tasks, but they are designed to become ever more astute as feedback is received. They strive for actions that are rewarded and then optimise themselves in response. They are effective in a scenario where little training data exists and there is no clearly defined end state. For example, balancing the load on electricity grids in varying demand and supply cycles, or the optimisation of self-driving cars.

A fourth machine learning technique is deep learning using neural networks. This technique has made impressive advances in image processing and recognition, robotics and natural language processing.

Are these advances achievable for industrial plants and industrial automation? The picture is more nuanced than the AI consultants would have us believe. There are examples of singular successes and commercially useful applications that have lured us into the belief that AI can be ably applied to industrial plants. In fact, an almost superstitious awe exists around the capacity of AI.

The widely quoted industrial IoT guru, Jonas Berge¹, suggests caution. For instance, a dubious claim is that an analysis of existing plant data (such as the correlations between flow data and pump failures) will provide new insights and ultimately optimise plant processes. Often this costly exercise merely reveals what you already knew or suspected.

Rather than complex machine learning, a simple rules-based system is best when analysing plant data. A flow sensor, for instance, is level-sensitive; it acts by setting off an alarm or shutting down a pump, or closing a valve.

It has been suggested that AI is best placed to analyse humans; to find useful patterns from a morass of complexity. Machines, on the other hand, are predictable: it is harder to gain value from applying AI and machine learning to machines.

In conclusion, instrument specialists remain critical to industrial plants; their deep knowledge of the processes and operations ensure they run smoothly. Yes, AI is producing innovations (Chao and You 2019²), but not widely yet in terms of having a genuine and useful impact on plants.

*Steve Mackay PhD has worked across the world for the past 40 years in the design and construction of iron ore plants, oil and gas platforms and power stations, as well as plant maintenance. He believes university engineering programs need to be strongly focused on industry. He has been the author or editor of over 30 engineering textbooks and is currently leading the first fully online accredited engineering college with over 1500 students from over 140 countries.

Reference

1. Berge J 2020, Big Ideas 2021: The New Data Science, LinkedIn, <<https://www.linkedin.com/pulse/bigideas2021-new-data-science-jonas-berge?trk=portfolio_article-card_title>>
2. Shang C and You F 2019, ‘Data Analytics and Machine Learning for Smart Process Manufacturing: Recent Advances and Perspectives in the Big Data Era’, Engineering, vol. 5, issue 6, pp101-1016.

Image: ©stock.adobe.com/au/faithie